Deep neural networks (DNNs) have recently emerged as an attractive solution for many learning problems. DNNs can offer great representational power, without demanding much feature engineering. Computational advances and the availability of large datasets have made it feasible to train large DNNs, with millions of parameters.